An entity-guided text summarization framework with relational heterogeneous graph neural network

TL;DR

This paper introduces a novel entity-guided text summarization framework that leverages relational heterogeneous graph neural networks and knowledge graphs to improve the faithfulness and relevance of summaries.

Contribution

It proposes a new framework combining GNNs and KGs with entity-guided multi-task learning for more effective text summarization.

Findings

Outperforms existing extractive and abstractive baselines on CNN/DM and NYT50 datasets.

Entity-entity edge density in the graph significantly impacts summarization performance.

Ablation studies confirm the effectiveness of the proposed components.

Abstract

Two crucial issues for text summarization to generate faithful summaries are to make use of knowledge beyond text and to make use of cross-sentence relations in text. Intuitive ways for the two issues are Knowledge Graph (KG) and Graph Neural Network (GNN) respectively. Entities are semantic units in text and in KG. This paper focuses on both issues by leveraging entities mentioned in text to connect GNN and KG for summarization. Firstly, entities are leveraged to construct a sentence-entity graph with weighted multi-type edges to model sentence relations, and a relational heterogeneous GNN for summarization is proposed to calculate node encodings. Secondly, entities are leveraged to link the graph to KG to collect knowledge. Thirdly, entities guide a two-step summarization framework defining a multi-task selector to select salient sentences and entities, and using an entity-focused abstractor to compress the sentences. GNN is connected with KG by constructing sentence-entity graphs where entity-entity edges are built based on KG, initializing entity embeddings on KG, and training entity embeddings using entity-entity edges. The relational heterogeneous GNN utilizes both edge weights and edge types in GNN to calculate graphs with weighted multi-type edges. Experiments show the proposed method outperforms extractive baselines including the HGNN-based HGNNSum and abstractive baselines including the entity-driven SENECA on CNN/DM, and outperforms most baselines on NYT50. Experiments on sub-datasets show the density of sentence-entity edges greatly influences the performance of the proposed method. The greater the density, the better the performance. Ablations show effectiveness of the method.